Skip to content

DeepTeam is a framework to red team LLMs and AI agents.

results printing

penguine2173cab35f1b
diff --git a/deepteam/attacks/multi_turn/linear_jailbreaking/__pycache__/linear_jailbreaking.cpython-311.pyc b/deepteam/attacks/multi_turn/linear_jailbreaking/__pycache__/linear_jailbreaking.cpython-311.pyc
index 07ae9822..b2ff400f 100644
Binary files a/deepteam/attacks/multi_turn/linear_jailbreaking/__pycache__/linear_jailbreaking.cpython-311.pyc and b/deepteam/attacks/multi_turn/linear_jailbreaking/__pycache__/linear_jailbreaking.cpython-311.pyc differ
diff --git a/deepteam/red_teamer/__pycache__/red_teamer.cpython-311.pyc b/deepteam/red_teamer/__pycache__/red_teamer.cpython-311.pyc
index 7c1b0a3f..f6555a4a 100644
Binary files a/deepteam/red_teamer/__pycache__/red_teamer.cpython-311.pyc and b/deepteam/red_teamer/__pycache__/red_teamer.cpython-311.pyc differ
diff --git a/deepteam/red_teamer/__pycache__/risk_assessment.cpython-311.pyc b/deepteam/red_teamer/__pycache__/risk_assessment.cpython-311.pyc
index 0f0d68c0..0da1213e 100644
Binary files a/deepteam/red_teamer/__pycache__/risk_assessment.cpython-311.pyc and b/deepteam/red_teamer/__pycache__/risk_assessment.cpython-311.pyc differ
diff --git a/deepteam/red_teamer/red_teamer.py b/deepteam/red_teamer/red_teamer.py
index e8a5f41a..16692069 100644
--- a/deepteam/red_teamer/red_teamer.py
+++ b/deepteam/red_teamer/red_teamer.py
@@ -1,7 +1,7 @@
import asyncio
from tqdm import tqdm
from typing import Dict, List, Optional, Union
-from pydantic import BaseModel
+from rich.console import Console
import inspect
@@ -219,6 +219,7 @@ class RedTeamer:
self.save_test_cases_as_simulated_attacks(
test_cases=red_teaming_test_cases
)
+ self._print_risk_assessment()
return self.risk_assessment
async def a_red_team(
@@ -323,6 +324,7 @@ class RedTeamer:
self.save_test_cases_as_simulated_attacks(
test_cases=red_teaming_test_cases
)
+ self._print_risk_assessment()
return self.risk_assessment
async def _a_attack(
@@ -563,3 +565,71 @@ class RedTeamer:
simulated_attacks.append(simulated_attack)
self.simulated_attacks = simulated_attacks
+
+ def _print_risk_assessment(self):
+ if self.risk_assessment is None:
+ return
+
+ console = Console()
+
+ console.print("\n" + "=" * 80)
+ console.print(
+ "[bold magenta]πŸ” DeepTeam Risk Assessment[/bold magenta]"
+ )
+ console.print("=" * 80)
+
+ # Sort vulnerability type results by pass rate in descending order
+ sorted_vulnerability_results = sorted(
+ self.risk_assessment.overview.vulnerability_type_results,
+ key=lambda x: x.pass_rate,
+ reverse=True,
+ )
+
+ # Print overview summary
+ console.print(
+ f"\n⚠️ Overview by Vulnerabilities ({len(sorted_vulnerability_results)})"
+ )
+ console.print("-" * 80)
+
+ # Convert vulnerability type results to a table format
+ for result in sorted_vulnerability_results:
+ if result.pass_rate >= 0.8:
+ status = "[rgb(5,245,141)]βœ“ PASS[/rgb(5,245,141)]"
+ elif result.pass_rate >= 0.5:
+ status = "[rgb(255,171,0)]⚠ WARNING[/rgb(255,171,0)]"
+ else:
+ status = "[rgb(255,85,85)]βœ— FAIL[/rgb(255,85,85)]"
+
+ console.print(
+ f"{status} | {result.vulnerability} ({result.vulnerability_type.value}) | Mitigation Rate: {result.pass_rate:.2%} ({result.passing}/{result.passing + result.failing})"
+ )
+
+ # Sort attack method results by pass rate in descending order
+ sorted_attack_method_results = sorted(
+ self.risk_assessment.overview.attack_method_results,
+ key=lambda x: x.pass_rate,
+ reverse=True,
+ )
+
+ # Print attack methods overview
+ console.print(
+ f"\nπŸ’₯ Overview by Attack Methods ({len(sorted_attack_method_results)})"
+ )
+ console.print("-" * 80)
+
+ # Convert attack method results to a table format
+ for result in sorted_attack_method_results:
+ if result.pass_rate >= 0.8:
+ status = "[rgb(5,245,141)]βœ“ PASS[/rgb(5,245,141)]"
+ elif result.pass_rate >= 0.5:
+ status = "[rgb(255,171,0)]⚠ WARNING[/rgb(255,171,0)]"
+ else:
+ status = "[rgb(255,85,85)]βœ— FAIL[/rgb(255,85,85)]"
+
+ console.print(
+ f"{status} | {result.attack_method} | Mitigation Rate: {result.pass_rate:.2%} ({result.passing}/{result.passing + result.failing})"
+ )
+
+ console.print("\n" + "=" * 80)
+ console.print("[bold magenta]LLM red teaming complete.[/bold magenta]")
+ console.print("=" * 80 + "\n")
diff --git a/deepteam/red_teamer/risk_assessment.py b/deepteam/red_teamer/risk_assessment.py
index 36f5a9f1..daa582de 100644
--- a/deepteam/red_teamer/risk_assessment.py
+++ b/deepteam/red_teamer/risk_assessment.py
@@ -8,15 +8,6 @@ from enum import Enum
from deepteam.vulnerabilities.types import VulnerabilityType
-class VulnerabilityTypeResult(BaseModel):
- vulnerability: str
- vulnerability_type: VulnerabilityType
- pass_rate: float
- passing: int
- failing: int
- errored: int
-
-
class RedTeamingTestCase(BaseModel):
vulnerability: str
vulnerability_type: VulnerabilityType
@@ -58,8 +49,26 @@ class TestCasesList(list):
return pd.DataFrame(data)
+class VulnerabilityTypeResult(BaseModel):
+ vulnerability: str
+ vulnerability_type: VulnerabilityType
+ pass_rate: float
+ passing: int
+ failing: int
+ errored: int
+
+
+class AttackMethodResult(BaseModel):
+ attack_method: str
+ pass_rate: float
+ passing: int
+ failing: int
+ errored: int
+
+
class RedTeamingOverview(BaseModel):
vulnerability_type_results: List[VulnerabilityTypeResult]
+ attack_method_results: List[AttackMethodResult]
def to_df(self):
import pandas as pd
@@ -138,6 +147,7 @@ def construct_risk_assessment_overview(
# Calculate statistics for each vulnerability type
vulnerability_type_results = []
+ attack_method_results = []
for vuln_type, test_cases in vulnerability_type_to_cases.items():
# Count passing, failing, and errored cases
@@ -160,8 +170,8 @@ def construct_risk_assessment_overview(
# Use the vulnerability name from the first case in this group
vulnerability_name = test_cases[0].vulnerability if test_cases else ""
- # Create the result object
- result = VulnerabilityTypeResult(
+ # Create the result objects
+ vulnerability_type_result = VulnerabilityTypeResult(
vulnerability=vulnerability_name,
vulnerability_type=vuln_type,
pass_rate=pass_rate,
@@ -169,10 +179,19 @@ def construct_risk_assessment_overview(
failing=failing,
errored=errored,
)
+ attack_method_result = AttackMethodResult(
+ attack_method=test_case.attack_method,
+ pass_rate=pass_rate,
+ passing=passing,
+ failing=failing,
+ errored=errored,
+ )
- vulnerability_type_results.append(result)
+ vulnerability_type_results.append(vulnerability_type_result)
+ attack_method_results.append(attack_method_result)
# Create and return the final assessment
return RedTeamingOverview(
- vulnerability_type_results=vulnerability_type_results
+ vulnerability_type_results=vulnerability_type_results,
+ attack_method_results=attack_method_results,
)